Reflexio Compiles Your Agent's Mistakes Into Rules
Reflexio (https://www.reflexio.ai) took #2 on Product Hunt with 282 upvotes, and it earns a look because it attacks the memory problem from the right end. It watches your agent in production, and every user correction, task failure, and success gets transformed into a behavioral rule the agent applies going forward. Not transcript storage with retrieval on top β compiled lessons. Conflicting feedback is handled with a sensible hierarchy: learn user-specific improvements first, then roll genuinely common patterns into rules that generalize across users.
The claimed numbers from case studies: 36% reduction in task-failure rate, 57% reduction in token usage, quality improvements in 47% of interactions with negligible regressions. Python SDK, REST API and CLI today, TypeScript coming, with ready-made skills for LangGraph and CrewAI. Founders are Yi Lu, a former Meta tech lead, and Guangyu Yang. Free tier plus 30 days of Pro at signup.
The detail that actually separates it: "every learning visible, testable, and reversible," with export and delete under user control. Most memory products bolt a vector store onto an agent and call it learning; you can't inspect what was learned, test whether a given lesson helps, or revert the one that's quietly making things worse. Making each learned rule a first-class object you can audit is the version of agent memory that survives contact with a compliance team β and the memory layer keeps being the most-read topic on this site for exactly that reason. It's the layer of the stack nobody has locked up yet.
← Back to all articles
The claimed numbers from case studies: 36% reduction in task-failure rate, 57% reduction in token usage, quality improvements in 47% of interactions with negligible regressions. Python SDK, REST API and CLI today, TypeScript coming, with ready-made skills for LangGraph and CrewAI. Founders are Yi Lu, a former Meta tech lead, and Guangyu Yang. Free tier plus 30 days of Pro at signup.
The detail that actually separates it: "every learning visible, testable, and reversible," with export and delete under user control. Most memory products bolt a vector store onto an agent and call it learning; you can't inspect what was learned, test whether a given lesson helps, or revert the one that's quietly making things worse. Making each learned rule a first-class object you can audit is the version of agent memory that survives contact with a compliance team β and the memory layer keeps being the most-read topic on this site for exactly that reason. It's the layer of the stack nobody has locked up yet.
Comments